Microsoft has announced a series of infrastructure updates during its Microsoft Build 2026 conference, including that its latest generation of Cobalt Arm chips is now available in early access preview.

The Cobalt 200 was detailed back in November 2025 by Microsoft, and according to the company, it has a 50 percent increase in performance compared to the previous generation, Cobalt 100, and is the result of an evaluation of more than 350,000 configuration candidates.

Each Cobalt 200 SoC includes 132 active cores with 3MB of L2 cache per core and 192MB of L3 system cache. According to Microsoft, the Cobalt CPUs also offer individual per-core dynamic voltage and frequency scaling, which enables each of the 132 cores to run at a different performance level.

Azure Cobalt 200 Arm-based Virtual Machines (VMs) are now available in early access preview. The VMs can scale up to 128 vCPUs. According to Microsoft, compared to the previous generation, they offer up to 135 percent better performance for cloud database workloads and 80 percent for caching workloads, and are fully compatible with workloads currently running on the previous generation chip.

The new VMs will be available in preview in West US3, East US2, Central US, Sweden Central, East US, West US2, Spain Central, and Indonesia Central, with additional regions to be announced.

Also during the Build 2026 event, Microsoft revealed that its Maia 200 AI accelerator - an AI inferencing chip revealed in January 2026 - is now available in production in Iowa and Arizona, with plans to expand to Italy, Australia, and South Korea next.

The Maia 200 is built using TSMC's 3nm technology and delivers around 10 petaflops of FP4 and five petaflops of FP8 compute performance within a 750W SoC (System-on-Chip) TDP (Thermal Design Power) envelope. The chip has a “redesigned memory system” comprising 216GB HBM3e at seven Tbps and 272MB of on-chip SRAM, alongside data movement engines.

This, claims Microsoft, makes the Maia 200 the most performant chip from any hyperscaler, offering three times the FP4 performance of Amazon’s Tranium3 and Google’s seventh-generation TPU, Ironwood.